Kodelet
Engineering capabilities, amplified

When Your AI Agent Just Goes Off and Does It

Published on by Jingkai He

Imagine you pose a casual question to a colleague about their approach to something - not expecting them to actually do it, just wanting to understand the method. A few minutes later, they return with the task not only completed but with a clear walkthrough of their process. Your reaction? It might depend on the person and the context, but I’d argue this behaviour is the hallmark of a high-agency top performer — someone who sees a problem and instinctively moves to solve it.

This is what unfolded in the story I’m about to tell, except my “colleague” was an AI agent powered by LLM that didn’t just understand the task at hand, but took initiative and simply got on with it.

It started with a casual question

This is what roughly happened: On a Sunday evening, I was sitting on the sofa, bored by Squid Game, SSH’d into my remote VM through Termius on my iPhone, and I wanted to set up gVisor with Docker for some extra sandboxing. Instead of diving into documentation or crafting the perfect installation steps, I casually typed:

kodelet run how can i install gvisor with docker on this machinr

Notice the typo? Yeah, “machinr” instead of “machine.” Did it matter? Not one bit.

What I expected vs. what happened

Here’s what I thought would happen: kodelet would explain the process, maybe give me some commands to run, and I’d carefully execute them step by step.

Here’s what actually happened: kodelet took one look at my question and said, “Hold my beer.”

It just… went off and did it. No questions asked. No “would you like me to proceed?” It carried out the following steps with some short and concise explanations in between:

  1. Analysed my system - checked that I was on Ubuntu 24.04 with Docker already installed
  2. Added the gVisor repository - GPG keys, apt sources, the works
  3. Installed runsc - the gVisor runtime component
  4. Configured Docker - updated the daemon configuration to recognise the new runtime
  5. Restarted Docker - because that’s what you do after configuration changes
  6. Checked the Docker runtime - ran docker info | grep -A 10 "Runtimes" to verify runsc’s presence
  7. Smoke tested everything - ran a hello-world container with the new runtime to prove it worked

The whole thing took maybe a minute. I sat there watching my terminal scroll by, completely in awe of the efficiency.

Wait… I didn’t ask for it?

This is where it gets interesting from a behaviour perspective. Many people would look at this and think, “The LLM didn’t follow instructions! I asked HOW to do it, not to DO it!”

But here’s the thing - this is exactly what you’d expect from a high-agency top performer. When you ask a seasoned DevOps engineer “how do I install gVisor?”, they don’t hand you a manual. They just do it and few minutes later come back to you with the result, because they can, and because doing is often faster than explaining.

This is the “You can just do things” mentality that separates good engineers from great ones. LLM demonstrated that same principle.

The follow-up: documentation and sharing

Of course, being someone who likes to understand what just happened, I followed up:

kodelet run -f writedown the instruction in gvisor.md

The -f flag means “follow up on the previous conversation” - one of my favourite kodelet features. Within seconds, I had a comprehensive gvisor.md file with step-by-step instructions, explanations of what gVisor is, usage examples, and troubleshooting tips.

Then, because I’m a bit of a digital packrat who likes to keep useful things accessible, I asked:

kodelet run -f upload it as a private gist using gh command

This is where things got even more interesting. kodelet wasn’t 100% sure about the exact syntax for creating private gists, so it did what any good engineer would do: it ran gh gist create --help to figure out the command structure, then executed it successfully.

That kind of adaptive problem-solving - “I don’t know the exact syntax, let me check the help and figure it out” - is exactly the behaviour you want from a capable assistant.

Reflections on speed vs. understanding

Here’s something that struck me: if I had done this installation myself, it probably would have taken 20-30 minutes. I’m fairly seasoned with DevOps work, so it’s not like I would have struggled. But watching kodelet do it was not only faster - it was educational.

I learned from the process even though I didn’t do the work myself. I saw the exact sequence of commands, understood the reasoning behind each step, and now I have documentation I can reference later.

There’s something to be said for this approach to learning: sometimes the best way to understand a process is to watch an expert execute it flawlessly, then ask them to explain what they did.

The most amusing part of this whole experience is that I probably spent more time writing this blog post about the kodelet interaction than kodelet spent actually solving the original problem.

But that’s the point, isn’t it? The novelty of autonomous AI agent operation is still growing on me. We’re living through a moment where you can casually ask an AI to configure your system infrastructure from your phone while sitting in a coffee shop, and it just… works.

Final thoughts: the comeback of DevOps

AI agents aren’t just about getting tasks done faster, they’re about amplifying human capability. kodelet freed me up to think about higher-level questions like “Now that I have gVisor running, how can I build useful stuff on top of it?” rather than wrestling with installation mechanics.

This makes me think: the industry moved away from DevOps towards platform engineering largely because embedding infra specialists in every team became unsustainable post-ZIRP due to the economic downturn. The “internal customer support” burden on SRE teams grew ever bigger.

Now with AI agents, instead of stretching human specialists across teams, you could democratise the knowledge and specialities by having AI agents handling 90% of routine problems at a fraction of the cost. This could massively reduce the support burden on centralised teams, letting them focus on strategic, organisation-wide services.

For me, watching agent doing system administration work while I lounged with my iPhone felt like a glimpse into a future where DevOps might just make its comeback, powered by LLM.

And frankly, I’m excited to see where this leads.

Want to try kodelet yourself? Check out the GitHub repository or visit kodelet.com for more information.